JPH07306149A - Method for discriminating agglomerating pattern of particles using neural network - Google Patents

Abstract

PURPOSE:To highly reliably measure the agglutination of particles based on the output signal of a neural network by generating a two-dimensional image signal by scanning an inclined base, dividing the base into a plurality of areas, integrating the intensity of light in each area, and inputting the integrated intensity of light to the neural network as the intensity of light in each area. CONSTITUTION:The images of all wells 2 are caught by using a solid-state image sensor 3, lens 4, and lighting lamp by moving a micro-plate 1 in the directions A and B by means of a motor 11. Taken image signals are supplied to an A/D converter 7 through an amplifier 6 and converted signals are stored in an image storing device 8 as two-dintensional image signals. A host processor unit 14 digitally processes the image signals. The two dimensional images of the holes are divided into, for example, 45 concentric rings and the average intensity of light of each ring is found. The found average intensity of light is inputted to a neural network built in the processor unit 14 as a feature. Therefore, the presence/absence of agglomeration can be accurately discriminated based on the output signal of the neural network.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for detecting and determining a particle aggregation pattern, and this method is used for diagnosis.

[0002]

BACKGROUND OF THE INVENTION Devices have been developed for analyzing blood samples by detecting the pattern of particle agglomeration formed on a reaction vessel having a sloping bottom.
Such a device is described, for example, on February 23, 1988 by K. Hijika.
It is described in US Pat. No. 4,727,033 issued to ta et al. In this device, a test liquid is formed in a conical reaction vessel by supplying a blood sample (that is, a blood sample or a serum sample) and a reagent (serum reagent or sensitized particle reagent) into the reaction vessel. The reaction container containing the test solution is allowed to stand for a predetermined time (for example, 30 minutes). During this reaction time, the particles in the test solution settle on the inclined bottom surface. When the particles agglomerate, they form a uniform agglomeration layer on the bottom surface. However, when the particles do not agglomerate with each other, they roll down along the sloping bottom surface and collect in the lowest central portion of the bottom surface to form dots. Next, the particle pattern formed on the bottom surface of the reaction container is detected by projecting light from one side of the reaction container and receiving the light transmitted through the reaction container by the photodetector.

To this end, US Pat. No. 4,727,033
JP-A No. 1994-242242 discloses a photodetector having two concentric light receiving regions. One is for receiving the light transmitted through the central part of the reaction container, and the other is for receiving the light transmitted through the peripheral part of the reaction container. By processing the output signals supplied from these light receiving regions, it is possible to determine whether the particle pattern is an agglomeration pattern or a non-aggregation pattern. When the particles are agglomerated, there is no noticeable difference between the output signals from the two light receiving areas. On the other hand, when the non-aggregated particle pattern is formed, the light transmitted through the central portion of the reaction vessel becomes weaker than the light transmitted through the peripheral portion, so that the output signal between the two light receiving regions is thereby reduced. The difference will be greater.

Therefore, the ratio of the output signal from the central light receiving area to the output signal from the peripheral light receiving area is obtained,
By comparing this ratio with predetermined upper and lower thresholds, the particle pattern formed on the inclined bottom surface can be determined and the blood sample analyzed. If the ratio is above the upper threshold, this indicates that aggregation has occurred. Also, if the ratio is less than the lower threshold, this indicates that no aggregation has occurred. If the measured ratio lies between the upper and lower thresholds, this indicates that the particle pattern cannot be reliably determined.

An improved method for determining particle patterns is disclosed in Japanese published patent application (JP 03-056843) corresponding to US patent application No. 384,497 filed June 25, 1989. This method consists of optoelectrically scanning the particle pattern to generate an image signal representing a two-dimensional image containing the entire particle pattern. The image signal is then processed to generate at least two features of the grain pattern, and the grain pattern is then determined based on these features.

One of the above characteristics is the ratio of the average output signal from the central portion of the image signal to the average output signal from the peripheral portion of the image signal. This is similar to the ratio used in the same thresholding technique according to US Pat. No. 4,727,033, supra. This two-dimensional image signal allows more particle patterns to be used in the ratio calculation, thereby improving decision accuracy. The second feature is
It is a measurement of the sharpness of dots in the middle and deep areas. If the sharpness of the mid-deep dots from the particle pattern is higher than a predetermined threshold, this indicates that no agglomeration has occurred.
If the sharpness of the central dot from the particle pattern is below a predetermined threshold, this indicates that agglomeration has occurred.
If these two characteristics, that is, the ratio and the sharpness are contradictory, it indicates that the particle pattern cannot be reliably determined.

The above two known methods for determining particle patterns are suitable for determining actual or typical particle patterns. However, in actual analysis, particle patterns are sometimes formed that cannot be clearly determined by either of the two known methods. In such a case, an automatic analysis result is generated indicating that the sample cannot be reliably determined, and the sample must be visually analyzed or retested. In order to improve the efficiency of the analysis, it is desirable to reduce the number of samples that need to be analyzed again. In two known methods, this is done by changing the threshold used for comparison, but the accuracy of the analysis is reduced by false positives.

[0008]

SUMMARY OF THE INVENTION The primary object of the present invention is to provide a reliable and economical method that maximizes the accuracy of the analysis and reduces the number of samples that require retesting or visual analysis. In a conventional manner, to provide a new and useful method for detecting and determining the particle pattern formed on the inclined bottom surface of the reaction vessel.

[0009]

The present invention achieves the above object by adopting a neural network as a method for determining a particle pattern.

The present invention is a method for measuring an agglutination reaction from a particle pattern formed on a slanted bottom surface of a reaction vessel, wherein the slanted bottom surface is photoelectrically scanned to obtain a two-dimensional particle pattern. Generating an image signal representing an image, dividing the inclined bottom surface into a plurality of areas due to different contours of the bottom surface, and integrating the light intensities in each area to generate an area light intensity According to the method, the image signal is processed into an area light intensity, the area light intensity is input to a neural network to generate an output signal, and the agglutination reaction is measured based on the output signal. Provide a way.

In the method according to the invention, said neural network is arranged such that said area light intensity of a predetermined number of particle patterns and the corresponding truth decision of a predetermined number of particle patterns are input to said neural network for training. I was trained. In this way, as the number of particle patterns input to the neural network for training increases, the number of samples that normally must be re-analyzed decreases and the accuracy of the analysis improves.

The details of the present invention will be described below.

FIG. 1 is a block diagram showing the configuration of one embodiment of an apparatus for carrying out the method of the present invention. In this embodiment, a microplate 1 in which many wells 2 functioning as reaction vessels are formed in a matrix is used. It should be noted that the method according to the present invention can be carried out using various reaction vessels other than the microplate.

Referring to FIG. 2, there is shown a cross-sectional view of the well 2 formed on the top surface of the microplate 1. The well 2 has a conical bottom surface, and on the bottom surface,
Many fine steps are uniformly formed to form a base layer of particles. This base layer is particularly suitable for forming a stable agglomerated particle pattern. In this aspect,
The well 2 has a diameter of 6 mm, and the microplate 1 is made of a transparent material such as transparent acrylic resin or glass.

In this embodiment, the image of the microplate 1 is a photoelectric image detector comprising a solid-state image sensor 3, a lens 4 arranged in front of the image sensor and an illumination lamp 5 arranged below the microplate 1. Is imaged by. The microplate is moved in the A and B directions so that the respective images of all reaction vessels can be captured. Conventional mechanisms 11, 12 for driving the microplate 1 are provided. In addition, a device (not shown) for supplying the microplate 1 onto the reaction line, a device (not shown) for dispensing the sample liquid and the reagent liquid into the wells 2, and the microplate for the reaction line. Various devices are provided, such as devices for moving along (not shown), all of which are well known in the art. For example, these devices are described in U.S. Pat.
It is shown in No. 4,727,033.

The device further comprises an amplifier 6 which supplies the image signal to an analog-to-digital converter 7, the converted digital image signal being stored in an image storage device 8. The solid-state image sensor 3 is driven by a drive circuit 9 controlled by a timing signal generator 10. The analog-to-digital converter 7 and the storage device 8 are also controlled by the timing signal generator 10. The microplate 1 is driven in the A direction and the B direction by a motor 11 supplied with energy by a motor driver 12. The illumination lamp 5, which is a fluorescent lamp, is excited by the lamp power supply 13.

A host processor unit 14 is provided to interoperate the above circuit elements. The image signal stored in the image storage device 8 is digitally processed by the host processor unit 14 to determine the particle pattern formed on the bottom of the reaction vessel 2. The operation of the host processor unit 14 is as described in detail below.

In order to determine a particle pattern by a neural network, it is advantageous to decompose a two-dimensional image of the particle pattern into a series of concentric rings.
In this example, it was empirically determined that 45 rings gave the desired result, but other ring numbers may be applied. The photoelectric detector is designed so that the innermost ring of the decomposed two-dimensional image represents the center of the reaction vessel,
It is located on the center of the conically inclined bottom of the reaction vessel. FIG. 3 is a plan view showing a series of concentric rings.
FIG. 2 is a view showing a cross section of the reaction container. In this embodiment, each ring has a width of 0.05 mm and the total width of all the rings is 4.5 mm. By calculating the average intensity of each ring, 45 features of the particle pattern representing the distribution from the center of the conically tilted bottom surface are formed. These 45 features of the particle pattern are
Input to neural network.

The neural network used in the present invention is a standard backpropagation network (b
ack-propagation network). Backpropagation neural networks are organized for pattern classification. In this embodiment, the neural network is built in the host processing unit 14, which is hardware (model name 4 available from GATEWAY 2000).
86/33) and a neural network software package (Neural Works available from Neuroll Ware Inc.). It can be trained with the analytical results as standard or control data. This neural network
Training by a human operator of his own discretion or an external teacher in the form of the naked eye is also useful. A general neural network can be used in the present invention. Other commercially available neural network software packages include ANSim (Scienc
e Applications International Corporation) and MacBrain (available from Neurix).

The neural network has inputs 0 and 1
It has to be configured between and. Therefore, the 45 features must be normalized between 0 and 1 so that the maximum feature value is 1 and the ratio of other feature values to the maximum feature value is maintained at the same ratio. I have to. This neural network was designed to include 45 inputs and 2 outputs. This output is also limited to values between 0 and 1. FIG. 4 is a block diagram illustrating the determination of particle patterns. Here, 45 features of the particle pattern image are extracted, and then a trained neural network is used to generate two output signals, which are used to determine the particle agglutination reaction.

FIGS. 5A, B-FIGS. 8A, B are photographs of sample particle patterns having 45 features shown in histograms of values from 0 to 1. The leftmost feature value represents the innermost first ring, and the rightmost feature value represents the outermost last ring. In particular, FIG. 5A is a photograph of a typical non-aggregated particle pattern with particles migrating to the center or bottom of the reaction vessel. FIG. 5B shows the obtained 45 feature values, where the value at the left end is close to 0 and the value at the right end is 1.
It has a sharp contrast for values close to. This typically shows a dark area in the center of the particle pattern due to the concentration of particles in the center of the reaction vessel.

FIG. 6A is a photograph of a typical agglomerated particle pattern, where the particles are evenly distributed on the bottom surface of the reaction vessel. The 45 feature values shown in FIG. 6B show a uniform distribution corresponding to the above.

7A and 8A are exemplary photographs of uncertain particle patterns, the corresponding 45 feature values of which are shown in FIGS. 7B and 8B, respectively. In general,
The uncertain particle pattern is any pattern in which it is not possible to determine whether the particles are aggregated or not aggregated.

In order to allow a neural network to determine a particle pattern by the above-described method using the 45 inputs and 2 outputs, first of all, according to the judgment of a human operator, a standard sample or a plurality of samples is used. You have to train the neural network. Neural networks consist of many standard samples, including agglutinated, non-aggregated and uncertain samples.
Alternatively, it is trained by presenting non-standard samples to the network. Each sample presented to the neural network is in the form of its 45 feature values and 2 outputs (output 1 and output 2). When an agglutinated standard sample is used, or when the human judgment on the sample is agglutinated, the two desired outputs are output 1 = 1 and output 2 = 0.
When a non-agglutinated standard sample is used, or when the human judgment on the sample is no agglutination, the two desired outputs are output 1 = 0 and output 2 = 1. When an uncertain standard sample is used, or when the human judgment about the sample is uncertain, the two desired outputs are output 1 = 0 and output 2 = 0.

Once the neural network has been trained, when new samples are presented, the pattern of these samples could be determined based on the generalizations made during the training.
When the neural network is called to make a decision, the 45 feature values of the sample are presented and the decision is made based on the two output values according to Table 1 below.

Table 1 Output 1 Output 2 Judgment 10 + (Positive) 0 1- (Negative) 0 0? Note: + means aggregation.

-Means non-aggregated.

[0028] Means uncertain.

There are other methods for determining the particle pattern using a predetermined threshold. This method is effective when output 1 and output 2 still take intermediate values between 0 and 1 even after training. For example, if the first output exceeds the predetermined threshold value, this particle pattern is determined to have aggregation. If the second output is above the threshold, then the particle pattern is determined to be non-aggregated. If both outputs are below the threshold, the particle pattern is determined to be uncertain. Both outputs cannot exceed the threshold at the same time. In this aspect, the predetermined threshold is
0.98, which has been experimentally determined to give the best results.

[0030]

EXAMPLES To determine the accuracy of the present invention in comparison with the technique described in Japanese Patent Application Laid-Open No. 03-056843 (herein referred to as Ratio Technique) and in comparison with a modification of the present invention. I did one experiment. FIG. 9 is a plan view of a different embodiment of the invention for comparison with the invention, where the decomposition of a two-dimensional image into a matrix of 49 squares is shown. The average intensity of each square forms the 49 features of the particle pattern and can be used with neural networks in a manner similar to the invention described above. FIG. 10 is a plan view showing a modification of the division method for comparison with the present invention. This shows the decomposition of the two-dimensional image of the particle pattern into eight equal wedge pies. The average intensity of each wedge forms 8 features of the particle pattern, which in addition to the 45 features of the present invention form 53 features, in the same manner as the previously described present invention, Used with networks. These eight pie wedges themselves
It cannot be trained for classification.

Each of the neural networks for the series of rings, the square matrix, the series of rings + paws, and the prior art ratio method have agglomerative, non-aggregative and ambiguous decisions with approximately equal distribution 67.
Trained with individual sample particle patterns. Each neural network and ratio method was then tested in the determination of 165 different samples, 79 of which were ambiguous. The results are shown in Table 2 below.

Table 2 Method Decision Accuracy Series of Rings 96% Ratio Method 89% Square Matrix 83% Series of Rings + Paw Edge 96% The above table shows that the series of rings is the most accurate of all techniques. It shows that it has Deformation with pie wedges gives comparable accuracy, but has the disadvantage of requiring more computation than the method of the invention. The present invention also provides
It is advantageous over the ratio method in that its accuracy is further improved. The accuracy of the invention can be further improved by taking the incorrectly determined sample according to the invention and adding it to the training set to retrain the neural network. The effect of the present invention is achieved by separating into a plurality of regions representing different heights of the inclined bottom surface. That is, by separating the sloping bottom into a number of regions representing different contours of the sloping bottom, it is possible to result in a division of the sloping bottom so that the image signal is processed to the area light intensity. .

For the square matrix embodiment of the invention, the accuracy of the decisions listed in Table 2 is only 83%, but it should be noted that with further training of the neural network, this result can be significantly improved. Is. In this regard, it can be seen that in this example only 67 particle pattern samples were used for training.

The present invention has also been compared with other techniques for computational efficiency. By using the present invention, measurement of samples can be achieved 10 times faster compared to the ratio method. This advantage can increase the throughput of a device that analyzes a large number of samples, or allows a more economical device with the same throughput. A square matrix has this advantage of the present invention, despite the 83% accuracy obtained using the square matrix embodiment.

FIG. 11 shows another embodiment of the neural network according to the present invention. This neural network has three outputs: output 1, output 2 and output 3. Otherwise, this embodiment is the same as the previously described embodiment.

In order for the neural network to determine the particle pattern, each sample has its 45 feature values and 3 outputs (output 1, output 2 and output 3).
Is presented to the neural network. When an agglutinated standard sample is used, or when the human determination of the sample is agglutinated, the desired output is
Output 1 = 0, output 2 = 0 and output 3 = 0. When a non-agglutinated standard sample is used, or when the human determination of the sample is no agglutination, the desired output is
Output 1 = 0, output 2 = 1 and output 3 = 0. When an uncertain standard sample is used, or when the source's determination of the sample is uncertain, the desired outputs are output 1 = 0, output 2 = 0 and output 3 = 1.

Once the neural network has been trained, when new samples are presented, the pattern of these samples could be determined based on the generalization made during the training.
When the neural network is called to make a decision, 45 feature values of the sample are presented and the decision is made based on the three output values according to Table 3 below.

Table 3 Output 1 Output 2 Output 3 Judgment 1 0 0 + (Positive) 0 1 0 − (Negative) 0 0 1? Note: + means aggregation.

-Means non-aggregated.

[0040] Means uncertain.

If the first output is the largest of the three outputs, the particle pattern is determined to be agglomerated. If the second output is the highest of the three outputs, the particle pattern is determined to be non-aggregated. If the third output is the largest of the three outputs, the particle pattern is determined to be unclear. The neural network of this modification has the same conditions as in the first embodiment described above in Table 2
Trained with a sample particle pattern. Then 165 different samples, 79 of which were ambiguous, were tested in this variant. The results, as in the first example, showed 96% accuracy for the series of ring decisions and 83% accuracy for the square matrix decisions.

As will be appreciated by those skilled in the art, the present invention is not limited to the embodiments described above, and various changes and modifications can be made without departing from the scope of the present invention. In particular, the features of the two-dimensional image of the particle pattern are not limited to those used in the preferred embodiment. For example, concentric rings that are not equally spaced may be used. When using a tilted reaction vessel instead of a conically tilted reaction vessel,
The concentric rings can be replaced by parallel lines to divide the features that represent the particle pattern.

[0043]

As described above in detail,

[Brief description of drawings]

FIG. 1 is a block diagram showing an apparatus for carrying out the method according to the invention.

FIG. 2 is a cross-sectional view showing wells formed in a microplate used in the device shown in FIG.

FIG. 3 is a series used to resolve the image signal of the particle pattern from the well shown in FIG. 2 into 45 light intensity areas representing different contours of the conically tilted reaction vessel. FIG. 3 is a plan view showing the ring of FIG.

FIG. 4 is a block diagram showing a judgment operation.

5A is a photograph showing an image of a non-aggregated particle pattern (Sample 1), and FIG. 5B is a photograph showing its area light intensity value.

6A is a photograph showing an image of an agglomerated particle pattern (Sample 2), and FIG. 6B is a photograph showing an area light intensity value thereof.

FIG. 7A shows an unclear particle pattern (Sample 3).
7B is a photograph showing the image of FIG. 7B, and FIG. 7B is a photograph showing the area light intensity value.

FIG. 8A is an unclear particle pattern (Sample 4).
8B is a photograph showing the image of FIG. 8B, and FIG. 8B is a photograph showing the area light intensity value.

FIG. 6 shows a square light intensity area (square light in
FIG. 9 is a plan view showing another embodiment of the present invention in the form of (tensity areas).

FIG. 10 is a plan view showing a modification of the pie edge light intensity area form for comparison with the present invention.

FIG. 11 is a block diagram showing another embodiment of the judgment operation of the present invention.

[Explanation of symbols]

1 ... Microplate, 2 ... Well, 3 ... Solid-state image sensor, 4 ... Lens, 5 ... Illumination lamp, 6 ... Amplifier, 7
… Analog-to-digital converter, 8… Image storage,
9 ... Driving circuit, 10 ... Timing signal generator, 11 ... Motor, 12 ... Motor driver, 13 ... Lamp power supply, 14 ...
Host processor unit

Claims (11)

[Claims] 1. A method for measuring an agglutination reaction from a particle pattern formed on an inclined bottom surface of a reaction vessel, wherein the inclined bottom surface is photoelectrically scanned to obtain a two-dimensional image of the particle pattern. Generating an image signal representing, by dividing the inclined bottom surface into a plurality of areas due to different contours of the bottom surface, and integrating the light intensity in each area to generate an area light intensity, Processing the image signal into an area light intensity, inputting the area light intensity into a neural network,
A method comprising the steps of generating an output signal and measuring the agglutination reaction based on the output signal. 2. The method of claim 1, wherein the area light intensities of a plurality of particle patterns and the actual measurements corresponding to the plurality of particle patterns are input to the neural network for learning. The method further comprising the step of training the neural network. 3. The method according to claim 1, wherein two output signals are generated. 4. The method according to claim 1, wherein three output signals are generated. 5. The method according to claim 1, wherein the step of processing the image signal decomposes the image signal into a series of concentric rings, each ring representing a contour of a different area of the image. And measuring the average strength of each ring. 6. The method of claim 5, wherein the image signal is decomposed into 45 concentric rings. 7. A method for determining a particle pattern formed on a conically inclined bottom surface of a reaction vessel by using a neural network, wherein the particle pattern is photoelectrically scanned to obtain the particles. Generating an image signal representing a two-dimensional image containing a pattern; processing the image signal of the two-dimensional image to generate area light intensities representing different contours of the particle pattern; A method comprising: training with a sample; and determining a particle pattern with the trained neural network. 8. A method according to claim 7, wherein the step of processing the image signal comprises converting the image signal into a series of concentric rings, each ring representing a contour of a different area of the image. A method comprising disassembling and measuring the average strength of each ring. 9. The method according to claim 8, wherein the image signal is decomposed into 45 concentric rings. 10. The determination method according to claim 7, wherein:
The method wherein the determining step includes inputting the area light intensity to the neural network to generate an output signal, and measuring an agglutination reaction based on the output signal. 11. The determination method according to claim 7, wherein:
The method wherein the training step comprises inputting area light intensities of a plurality of particle patterns and actual measurements corresponding to the plurality of particle patterns into the neural network for learning.